Automatic stress analysis from facial videos based on deep facial action units recognition - s10044-021-01012-9.pdf

Tags

Top Highlights

• The extracted facial features are then fed to deep fully connected layers that regress AU intensities and robustly perform AU classification. The proposed algorithm is applied to the SRD’15 stress dataset, which contains neutral and stress states related to four types of stressors

• Apart from the AU classi- fication itself, the proposed AU classification network offers also a reliable confidence score about the presence of each AU in every input imag

• We use AUs for stress analysis and recognition because the AUs present different patterns between neutral and stress states. In most cases, the AU’s intensity is signifi- cantly higher during stress conditions as compared to a neu- tral state indicating a more expressive face

• Fig. 3 visualizes typical timeseries of an AU (AU17) for a neutral and a stressful task (adventure video)

• AU06, AU26, AU12, AU07, AU25, AU09, AU01, AU04, AU15, AU10, AU17, AU23, AU05, AU14

• AU06, AU26, AU17, AU25, AU10, AU23, AU12, AU15, AU07, AU05, AU14, AU01

• AU06, AU12, AU25, AU26, AU14, AU07, AU17, AU23, AU15, AU04, AU05, AU10

• An interesting conclusion is that stressful tasks lead to significant increased AU intensities, i.e. a more ”expressive” face.